Embedding physics in machine learning potentials

The last decade has seen an expansion in machine learning methods applied to atomistic modelling problems. Utilising the flexible functional form allowed by different flavours of machine learning approaches and thanks to the abundance of electronic structure data, it is now possible to fit highly predictive potential energy surfaces. However, transferability of these models is often limited, and predictive accuracy may only be expected in structural domains where the training data is concentrated. In the case of complex materials, where the accessible configurational space is significantly larger, data requirements could be prohibitive, and there is a need to place constrains on the interaction model. In this talk, I will discuss machine learning potentials in a Bayesian framework, and how physical knowledge and intuition can be embedded in the Bayesian prior of the model.